Daily Digest — Sep 23

TCMaster: Confidence-Aware Querying and Workload-Guided Physical Design for Multi-Source Traditional Chinese Medicine Knowledge Graphs

• TCMaster integrates pharmacopoeias, prescriptions, molecular databases, and LLM-extracted micro-semantics into a Traditional Chinese Medicine knowledge graph with approximately 221K entities and 723K base edges, annotating edges with provenance-level confidence.

Discovery-Driven Integration of Disjoint Tables via Text

• LOKI, a novel architecture, discovers fine-grained row-sentence associations between disjoint tables by learning contextualized representations and employing a global table-text contrastive objective. • LOKI outperforms state-of-the-art multi-modal data discovery approaches in integrating heterogeneous datasets and materializes typed integrated tables with 0.982 macro typed-pair precision.

ArticleMiner: Ontology-Guided Knowledge Graph Construction from Scientific Publications

• The ArticleMiner framework constructs knowledge graphs from scientific publications by integrating evidence from multiple parsers and a language model, guided by domain-specific task modules.

Semantic Abstraction for Natural Language Inference: a Methodological Framework for Discovering and Compensating Semantic Knowledge and Reasoning Gaps in Large Language Models

• A proposed methodological framework reconfigures lexical-semantic relations to construct higher-level abstract semantic knowledge, enhancing Large Language Models' (LLMs) natural language inference (NLI) capabilities. • This approach discovers and compensates for semantic knowledge gaps in LLMs, leading to over 10% accuracy improvement for some models, particularly in the non-entailment class.

DDGAD: Trajectory Dynamics for Diffusion-Based Graph Anomaly Detection

• DDGAD, a diffusion-based graph anomaly detection framework, utilizes trajectory dynamics to differentiate normal and anomalous nodes by observing representation stability under diffusion regularization and reliability-aware neighborhood consensus.

Skill-Conditioned Visual Geolocation for Vision-Language

• GeoSkill, a new training-free framework, introduces an evolving Skill-Graph for visual geolocation, addressing limitations in structured reasoning and self-evolution seen in existing Vision-Language Models (VLMs).

From Accuracy to Auditability: A Survey of Determinism in Financial AI Systems

• Nondeterminism in financial AI systems, especially deep neural networks and Generative AI, arises from hardware and architectural factors, impacting reproducibility in areas like credit risk, fraud detection, and anti-money laundering.

A Comparative Analysis of LLM Memorization at Statistical and Internal Levels: Cross-Model Commonalities and Model-Specific Signatures

• Memorization rate in LLMs scales log-linearly with model size, and memorized sequences exhibit compressibility and shared frequency/domain distribution patterns across different model families.